Content uploaded by Christos Papadelis
Author content
All content in this area was uploaded by Christos Papadelis on Feb 08, 2021
Content may be subject to copyright.
This article was published in an Elsevier journal. The attached copy
is furnished to the author for non-commercial research and
education use, including for instruction at the author’s institution,
sharing with colleagues and providing to institution administration.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
Author's personal copy
Monitoring sleepiness with on-board electrophysiological recordings
for preventing sleep-deprived traffic accidents
Christos Papadelis
a,*
, Zhe Chen
a,b
, Chrysoula Kourtidou-Papadeli
c
,
Panagiotis D. Bamidis
a
, Ioanna Chouvarda
a
, Evangelos Bekiaris
d
, Nikos Maglaveras
a
a
Aristotle University of Thessaloniki, School of Medicine, Laboratory of Medical Informatics, P.O. Box 323, 54124 Thessaloniki, Greece
b
RIKEN Brain Science Institute, Laboratory for Advanced Brain Signal Processing, Wako, Saitama, Japan
c
Greek Aerospace Medical Association and Space Research, Thessaloniki, Greece
d
Center for Research and Technology, Hellenic Institute of Transport, Thessaloniki, Greece
See Editorial, pages 1899–1900
Abstract
Objective: The objective of this study is the development and evaluation of efficient neurophysiological signal statistics, which may assess
the driver’s alertness level and serve as potential indicators of sleepiness in the design of an on-board countermeasure system.
Methods: Multichannel EEG, EOG, EMG, and ECG were recorded from sleep-deprived subjects exposed to real field driving condi-
tions. A number of severe driving errors occurred during the experiments. The analysis was performed in two main dimensions: the mac-
roscopic analysis that estimates the on-going temporal evolution of physiological measurements during the driving task, and the
microscopic event analysis that focuses on the physiological measurements’ alterations just before, during, and after the driving errors.
Two independent neurophysiologists visually interpreted the measurements. The EEG data were analyzed by using both linear and non-
linear analysis tools.
Results: We observed the occurrence of brief paroxysmal bursts of alpha activity and an increased synchrony among EEG channels
before the driving errors. The alpha relative band ratio (RBR) significantly increased, and the Cross Approximate Entropy that quan-
tifies the synchrony among channels also significantly decreased before the driving errors. Quantitative EEG analysis revealed significant
variations of RBR by driving time in the frequency bands of delta, alpha, beta, and gamma. Most of the estimated EEG statistics, such as
the Shannon Entropy, Kullback–Leibler Entropy, Coherence, and Cross-Approximate Entropy, were significantly affected by driving
time. We also observed an alteration of eyes blinking duration by increased driving time and a significant increase of eye blinks’ number
and duration before driving errors.
Conclusions: EEG and EOG are promising neurophysiological indicators of driver sleepiness and have the potential of monitoring sleepi-
ness in occupational settings incorporated in a sleepiness countermeasure device.
Significance: The occurrence of brief paroxysmal bursts of alpha activity before severe driving errors is described in detail for the first
time. Clear evidence is presented that eye-blinking statistics are sensitive to the driver’s sleepiness and should be considered in the design
of an efficient and driver-friendly sleepiness detection countermeasure device.
Ó2007 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.
Keywords: Sleepiness; Fatigue; Driving; EEG; EOG; Eye blinks
1. Introduction
Sleepiness at the wheel has been identified as the reason
behind fatal crashes and highway accidents caused by car
or truck drivers (Philip, 2005; Connor et al., 2001; Hakka-
nen and Summala, 2000). Boredom, fatigue, monotony,
1388-2457/$32.00 Ó2007 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.
doi:10.1016/j.clinph.2007.04.031
*
Corresponding author. Tel.: +30 2310 999332; fax: +30 2310 435331.
E-mail addresses: cpapad@med.auth.gr,papadeli@koz.forthnet.gr
(C. Papadelis).
www.elsevier.com/locate/clinph
Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
disturbed or deprived sleep may induce sleepiness and
drowsiness. Deprived sleep is one of the most important
factors of sleepiness that affects various aspects of perfor-
mance (Bocca and Denise, 2006). Sleep deprivation can
reduce attention and vigilance by 50%, decision-making
ability, communication skills, and memory (Killgore
et al., 2006; Raidy and Scharff, 2005; Harrison and Horne,
2000). The most sensitive tasks are those that are long,
monotonous, and boring such as driving (especially during
night) that becomes very vulnerable to the effects of sleep
deprivation (see Dinges and Kribbs, 1991, for a review).
During sleepy conditions, decreased attention, impaired
information processing, and the reduced decision-making
capability can all diminish driver’s ability to respond effec-
tively to unusual or emergent situations (Mascord and
Heath, 1992). Specifically, Williamson and Feyer (2000)
affirmed that sleep-deprived drivers are just as dangerous
as drunk drivers. Recent studies presented evidence that
driver sleepiness accounts for 6% of crashes, 15% of fatal
crashes, and 30% of fatal crashes on rural roads (The Par-
liament of the Commonwealth of Australia, 2000). Since
sleepiness impairs cognitive skills and therefore can
adversely affect drivers’ ability to monitor and assess their
own fitness to continue driving safely, serious care should
be taken for the implementation of sleepiness technological
countermeasures, which might be used to provide drivers
with useful feedback about the onset of sleepiness and to
improve road safety.
The importance of developing driver sleepiness counter-
measure devices has been identified in recent studies,
mainly for the purpose of preventing driving accidents
and errors (Lal and Craig, 2001b). The basic idea behind
vehicle-based detection is to monitor the driver unobtru-
sively by means of an on-board system that can detect
when the driver is impaired by sleepiness and drowsiness.
Such a device might possibly be based on physiological
measurements which are sensitive to the driver’s alertness.
Numerous physiological indicators are possible to assess
the sleepiness and alertness level. The electroencephalo-
graphic (EEG) signal may be one of the most predictive
and reliable measurements since it reflects directly human
brain activity (Artaud et al., 1973; Volow and Erwin,
1973).
Driver sleepiness research has received increased interest
in the last few decades. A number of studies have been per-
formed in drivers concerning the EEG alterations due to
sleepiness. A review of this literature can be found in Lal
and Craig (2001b). In most studies, the EEG data were
subjected to Fourier spectral analysis and alterations in
alpha and theta bands were generally reported (Torsvall
and A
˚kerstedt, 1983; A
˚kerstedt et al., 1982; Lal and Craig,
2000). In a pioneer study of night driving, sleep intruded
while the drivers still had their eyes open, and it was
accompanied by theta waves, sleep bursts and K-complexes
(O’Hanlon and Kelley, 1977). Interestingly, the drivers had
not been aware that they had been driving the car while
asleep. Later, Torsvall and A
˚kerstedt (1987) recorded
ambulatory EEG and electrooculogram (EOG) from train
drivers during a night and day trip and they observed that
alpha power increased during night driving relative to the
daytime levels for the sleepy group. Higher alpha power
during night trips was also observed in this sleepy group
in comparison with the alert group (Torsvall and A
˚ker-
stedt, 1987). Later, in another study of night driving, Lal
and Craig (2000) found, in a laboratory-based simulator
setup, consistent increases in delta, theta, and alpha activ-
ities during transition to sleepiness from an awake and alert
state. These authors focused their attention on the appear-
ance of delta waves during sleepiness and confirmed that
delta rhythm may become a reliable and simple indicator
of sleepiness.
A few authors have attempted further to establish a link
between the alertness and performance by correlating EEG
alterations and the driving task’s performance. Specifically,
Khardi and Vallet (1994) presented a significant positive
correlation between the number of steering wheel reversals
and the EEG’s activities in the theta and alpha bands.
More recently, Horne and Baulk (2004) established corre-
lations between the (alpha + theta) EEG power and the
incidents (a car wheel crossing the lateral lane marking)
in a simulated driving task. In agreement with these results,
Campagne et al. (2004) also established correlations
between the lower-frequency EEG changes and the driving
errors. In the independent driving simulator studies, it was
reported that the driving duration but not sleep deprivation
was found to have an effect on (alpha + theta) EEG power
(Otmani et al., 2005), or that the driving performance was
positively correlated with log sub-band (<20 Hz) power
spectrum (Lin et al., 2005). However, all these results have
been obtained using driving simulators. Consequently, sub-
jects knew that the consequences of their driving errors
would not affect their safety. Thus, from a psychological
point of view, it is necessary to perform further studies with
experimental paradigms in more realistic or field
environments.
The main goal of our study is the development of a reli-
able and driver-friendly neurophysiological fatigue coun-
termeasure device that may detect sleepiness, and that
can provide drivers with useful feedback and alert them
about the onset of sleepiness and consequently may assist
in the prevention of sleep-deprived car accidents. In our
study, EEG and peripheral physiological measurements
were collected from healthy sleep-deprived subjects during
realistic night driving conditions in an experimental car on
the road. The data analysis was performed in two main
dimensions: (a) macroscopic analysis, which monitors or
estimates the on-going temporal evolution of physiological
measurements during the driving task, assuming that sleep
deprivation together with the duration of driving would
attenuate gradually the driving performance as well as
the subjective sleepiness; and (b) microscopic event analy-
sis, which focuses on physiological measurements’ altera-
tions before and during driving incidents or serious
driving errors. The EEG data were analyzed by using both
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1907
Author's personal copy
standard linear techniques and the more advanced non-lin-
ear analysis tools.
2. Materials and methods
2.1. Subjects
Twenty-one subjects (20 males and 1 female) partici-
pated in the present study with a mean age of
33.04 ± 10.7 (mean ± SD) years. The subjects had average
driving experience (i.e., with possession of their driving
license) for 12.28 ± 8.66 (mean ± SD) years. All the drivers
reported that they drove more than 10,000 km/year. They
were medically evaluated prior to the study and were found
to be in good physical condition. All subjects had normal
vision or corrected to normal (20/20). One subject was
excluded from the subject pool before the experiment due
to severe hypertension. Alcohol, caffeine, tea, and choco-
late consumption were prohibited for one day prior to
the measurements. All subjects were free of any kind of
medication, and did not possess any personal or family his-
tory of neuropsychiatric disorder. Each subject signed an
informed consent agreement prior to his/her participation
and completed a short questionnaire. All subjects were paid
for their participation. Written instructions explaining to
the subjects the complete experiment procedure were pro-
vided. This investigation has obtained an approval from
the Local Ethics Committee.
2.2. Experimental protocol
The experiments were performed at CERTH-HIT (Cen-
ter for Research and Technology, Hellenic Institute of
Transport) in Thessaloniki, Greece, from 6 June up to 27
July 2005. The participants were supervised to ensure that
they remained awake for at least 24 h prior to the experi-
ment, and then to arrive at the CERTH-HIT around
8.00 pm. Upon arrival and after passing the standard med-
ical examination, the subjects’ level of sleepiness was esti-
mated in a 20-min recording at rest position by using the
Maintenance of Wakefulness Test (MWT) (Doghramji
et al., 1997). The EEG, EOG, electromyogram (EMG),
and electrocardiogram (ECG) measurements were col-
lected during this stage. The subjects were asked to relax
in a lean-back armchair in a quite, dark room and try
not to sleep while keeping their eyes open during the mea-
surements. An experienced sleep medical doctor conducted
the on-line monitoring of the EEG traces, and estimated
subject’s sleepiness level according to the MWT scale (1–
20, with 1 being most sleepy and 20 being most awake).
Their sleep behavior was also scaled by using the Epworth
Sleepiness test such that the pathological sleep disorders
were excluded. All subjects were found free of any patho-
logical sleep disorder.
The on-board measurements were conducted in the
CERTH-HIT experimental car (LANCIAäThesis, 2.4
Emblema, 5 cylinders, 2446 cm
3
, 170 HP). The vehicle is
equipped with double support pedals that are accessible
to the driver’s instructor and are equipped with advanced
driver assistance systems and sensors, such as the Lane
Detection System (LDS), and the Eye Leads Sensors
(ELS). The LDS (CRF, FIAT Center of Investigations)
consists of a CCD camera and a processing unit that recog-
nizes lane borders and estimates the position and orienta-
tion of the experimental vehicle with respect to them
(with a temporal resolution of 1 s). The ELS system (Sie-
mens, Germany) detects the eye blinks and consists of a
CCIR camera with two near infrared lighting units that
enable night measurements (see the inset of Fig. 1), and a
personal computer supporting the ELS software. The
ELS system outputs the following data: timestamp in sec-
onds, eye blink duration of both eyes in millisecond, and
the per minute averaged blink duration (PERCLOS).
The subject was seated at the driver seat and the
attached electrodes were connected to an ambulatory mon-
itoring system (Brain Products GmbH, Germany) (Fig. 1).
Eight-channel EEG, four-channel EOG, EMG, and ECG
recordings were collected simultaneously during the exper-
iment. An experienced driving instructor was seated at the
co-driver’s seat. At the back a technician monitored the
functioning of the recording equipment, and a medical doc-
tor monitored the on-going EEG and the other physiolog-
ical measurements. A CCIR camera was placed behind the
wheel that also monitored the subjects. The camera display
was shown on a LCD screen in the back seat. The driving
data collected by the LDS system were used for marking
the driving error events. Only the driving errors identified
by the driver instructor and verified by the LDS system
were used for the final quantitative analysis.
The same route was followed during each measurement
from CERTH-Thessaloniki to Veria (approximate dis-
tance: 100 km) and then back to Thessaloniki. It is a
monotonous motorway with light traffic during the exper-
Fig. 1. One of the experimental subjects in front of the wheel wearing the
EEG cap before the on-road experiment. The inset of the figure shows the
Eye Leads Sensors (ELS) system, which consists of a CCIR camera and
two near infrared lighting units that enable the night measurements.
1908 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
iment’s time. Subjects were asked to drive at their own pace
while observing the usual driving rules that are to drive in
the right lane except for overtake and not to exceed the
motorway speed limit (in Greece: 120 km/h). They were
also required to drive as much as possible in the center of
the right lane and to avoid overtakes. The task was monot-
onous enough to promote hypovigilance. In four cases,
subjects’ sleepiness levels were very high during the driving
task, and the driver instructor stopped the measurements
after two consecutive severe sleepiness incidents (e.g., unin-
tentionally crossing the lane border). During the experi-
mental procedure, a total of nine severe driving errors
occurred. All severe errors occurred during the last
15-min of the experiment. Seven out of nine events were
observed during the last 10-min. There were also minor
driving error events during the task. These events were
observed only by the driver instructor and not confirmed
by the LDS system. We excluded these minor driving error
events from our analysis in order to ensure that these
observations were not just driver instructor’s subjective
observations avoiding a possible bias in our analysis.
2.3. Physiological recordings
Physiological recordings were measured during the pre-
experimental stage (20-min recordings in a dark room) and
during the actual experiment (measurements in the research
car on the motorway) by means of a digital ambulatory
data acquisition system (Brain Products, Inc.). An Elec-
tro-Cap connected to the recording device was used to col-
lect EEG data from positions Fp1, Fp2, C3, C4, P3, P4,
O1, and O2. The active sites on the scalp were referenced
to linked mastoids, and all impedances were maintained
less than 10 kX. The EOG signal was also monitored via
four Ag/AgCl electrodes positioned one above, and one
below the right eye, one at the outer canthus of the right
eye and one at the nasion. EOG channels were connected,
one for horizontal (HEOG) and one for vertical (VEOG)
eye-movements. Electrodes for heart activity were posi-
tioned on the sternum and the fifth intercostal space on
the left side of the body. The EMG signal was also col-
lected from two Ag/AgCl electrodes positioned on the chin.
A 50-Hz hardware notch filter was also applied to all mea-
surements to remove power interference. Sampling fre-
quency was 200 Hz for all channels. Two test
measurements were performed before the whole experi-
mental procedure in order to assure that there was no noise
contamination from the vehicle electrical or mechanical
parts.
2.4. Artifact rejection
The EEG recordings were first band-pass filtered (2nd
order Butterworth filter, low-pass filter cut-off frequency:
40 Hz, high-pass filter cut-off frequency: 0.5 Hz), and then
the Infomax Independent Component Analysis (ICA) algo-
rithm was used, in order to remove the artifacts from the
data (e.g., Jung et al., 2000). The analysis and extraction
of artifacts were performed off-line on a PC by means of
the EEGLAB software (Delorme and Makeig, 2004). The
ICA decomposition was performed on the entire multi-
channel waveforms (eight EEG channels plus two EOG
channels). After the separation of the independent compo-
nents, all the abnormal components such as eye blinks,
eye-movements, and muscle activity were eliminated.
Components contaminated by artifacts were rejected, and
the remaining components were mixed and projected back
onto the scalp-channels.
The peripheral physiological data were also band-
passed filtered (EOG: 2nd order Butterworth filter,
low-pass filter cut-off frequency: 13 Hz, high-pass filter
cut-off frequency: 1 Hz; ECG: 2nd order Butterworth
filter, low-pass filter cut-off frequency: 40 Hz, high-pass
filter cut-off frequency: 1 Hz; EMG: 2nd order Butter-
worth filter, low-pass filter cut-off frequency: 100 Hz,
high-pass filter cut-off frequency: 20 Hz). For two subjects
no EMG data could be taken because of the removal of
electrodes during the experiment.
2.5. Statistical data analysis
The quantitative analysis was performed on the artifact-
free EEG and peripheral physiological data in two main
dimensions: macroscopic and microscopic. At the macro-
scopic analysis, we estimate the on-going temporal evolu-
tion of physiological measurements during the driving
task, assuming that sleep deprivation together with the
increased duration of driving would attenuate the driving
performance and subjective sleepiness would increase. At
the microscopic event analysis, we focus on the physiologi-
cal measurements’ alterations before, during, and after the
driving errors. The analysis was performed on the recorded
data that satisfied the following two criteria: (a) the sub-
ject’s MWT score before the experiment was less than 10,
and (b) the subject managed to accomplish one hour driv-
ing task such that all subjects have comparably equal dura-
tion of recordings. The EEG traces were visually
interpreted by two independent sleep-specialized neurophy-
siologists who kept notes on the traces. The specialists were
not aware of the possible driving errors, or the subject’s
MWT score. The visual interpretation of the data revealed
the occurrence of high-amplitude ‘alpha bursts’ synchroni-
zation events before the driving errors, which will be ana-
lyzed in detail in the next section of the paper.
In the macroscopic event analysis, EEG recordings were
divided into one-second segments. For each channel, the
relative band ratio (RBR) of the classic EEG frequency
bands (delta: 0.5–4 Hz, theta: 4–8 Hz, alpha: 8–12 Hz, beta:
12–30 Hz, and gamma: 30–40 Hz) was calculated. The
RBR is estimated by the relative ratio of the power of spe-
cific frequency bands against the total frequency power,
and it is a unit-less value in the range between 0 and 1.
The Shannon Entropy (SE) (Appendix A), the Kullback–
Leibler (K–L) Relative Entropy (Appendix B), and the
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1909
Author's personal copy
Approximate Entropy (ApEn) (Appendix C) were also cal-
culated for each EEG segment and each channel. As refer-
ence segment for the K–L entropy, an EEG segment from
the first minute of each recording was used. These quanti-
tative EEG statistics were then averaged in 5-min periods.
Based on the observations of the two neurophysiologists
who visually analyzed the signals, we applied both linear
and non-linear statistical tools to our measurements in
order to detect and quantify the multichannel synchroniza-
tion of ‘alpha spindles’ events that were identified as crucial
indicators of sleepiness. To quantify the synchrony among
pairwise channels, we use the common linear coherence
measure and the non-linear Cross-ApEn (Appendix D).
Coherence was calculated for the classic EEG frequency
bands per one-second segment between all possible combi-
nations of the EEG channels. As an information-theoretic
measure, Cross-ApEn describes the pattern similarity
between two time series. It can be viewed as a generaliza-
tion of the non-linear dynamic measure of ApEn, which
is a statistic that quantifies the complexity (or irregularity)
of a single signal (Fusheng et al., 2001). We applied the
cross-ApEn with typical setup of m= 1 and r= 0.2 to
the EEG segments of one second for all pairwise combina-
tions of channels. These quantitative EEG statistics were
also averaged in 5-min periods.
The eye blinks were detected as peaks in the differenti-
ated EOG signals, and the number of eye blinks was also
calculated per minute (Hyoki et al., 1998). The per minute
averaged eye blink duration (PERCLOSE) was also calcu-
lated from the EOG signal.
We examine the MWT score variation as a function of
the driver’s age and driver’s experience by using one-way
analysis of variance (ANOVA). Possible correlations
between the MWT score and the EEG and eye blink statis-
tics were also examined (Pearson’s bivariate correlation
test).
For the macroscopic analysis, we have complete data,
available to perform statistical analysis from 10 subjects
(EEG data were either missing or were interrupted in four
subjects due to technical problems during recordings, and
six subjects did not satisfy the inclusion criteria of the pres-
ent study). One-way ANOVA was applied to the averaged
EEG statistics. The independent variables were different
time factors (5, 10, and 15-min). The Bonferroni test was
used for correction of multiple comparisons. Post hoc anal-
ysis was performed using the Tukey HSD test and the
means were considered significantly different when the
probability of error was less than or equal to 0.05. Corre-
lation between the averaged EEG quantitative statistics
and the MWT scores was calculated using the Pearson
coefficient correlation. A cut-off correlation coefficient of
0.632 was regarded as a sign of significance.
We also validated the ELS system’s ability to reliably
estimate the number of eye blinks per minute and the per
minute averaged eye blink duration (PERCLOSE) compar-
ing the system’s outputs with the estimated eye blink statis-
tics derived from the EOG signal.
In the microscopic event analysis, we averaged the cor-
responding EEG statistics every 5 s; 40 s before, and 10 s
after the nine severe driving incidents that occurred during
the experiments. Additionally, nine 5-s segments from the
first minute of each measurement (from those during which
severe driving errors occurred) served as control condi-
tions. We also averaged the eye blink statistics in every
minute right before the driving errors. As a baseline, we
selected one-minute segments from the first five minutes
of each measurement. One-way ANOVA was applied on
the averaged EEG and eye blink statistics. The independent
variable was the time factor. The Bonferroni test was also
used for correction of multiple comparisons. Post hoc anal-
ysis was performed using the Tukey HSD test.
For the definition of the exact onset time of a severe dri-
ver error, both LDS system and EMG activity were used as
markers. When a driver error event was noticed by the dri-
ver instructor and confirmed by the LDS system, then the
event was defined as severe driver’s error. In all cases, the
instructor corrected the driver’s error for safety reasons.
During this time, the driver was alerted by his/her mistake
and/or instructor’s intervention, leading to a noticeable
and measurable EMG activity (Fig. 2). The onset of the
driver’s error was defined according to the EMG activity
(increase in EMG RMS (in lV) by at least 25% in compar-
ison with the previous 200 ms time window). The EMG
patterns were estimated by computing muscles activation
level (RMS, root mean square) (Gentili et al., 2006; Papad-
elis et al., 2007).
In order to analyze the microscopic events of driving
accidents, we are also interested in monitoring the short-
range (1–2 s) phase synchrony between different EEG
channels. In the neuroscience literature, phase synchrony
is an important phenomenon that might have some rela-
tionships with attention, cognition, and memory (Lachaux
et al., 1999). In neuroelectric or biomagnetic recordings, a
good measure for quantifying the instantaneous phase
relationship between two signals is the so-called phase
locking value (PLV; Appendix E). Specifically, PLV has
some advantages over the common coherence measure
(i.e., spectral covariance) since coherence can be applied
only to stationary signals and it does not specifically
quantify phase relationship (Lachaux et al., 1999). In
addition, the PLV statistic calculated from Hilbert trans-
form is a non-linear measure compared with the linear
coherence measure; moreover, the PLV is based on the
instantaneous phase difference between two signals, while
the coherence does not explicitly use the phase informa-
tion. In this paper, no statistical test was applied to the
PLV measure, since we only want to use the PLV for
quantifying the specific duration (generally <1 s) of tem-
porally local, synchronized events. Application of such a
measure for predicting the synchronized event is the topic
of a separate research that will be reported elsewhere. In
this paper, we will show by visual illustration the distinc-
tion of the PLV statistic during, before, and after the syn-
chronized events.
1910 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
3. Results
3.1. Visual interpretation of EEG
Two independent sleep-specialized neurophysiologists
visually interpreted the measurements. The visual interpre-
tation of the continuous measurements of EEG, EMG, and
EOG revealed the occurrence of ‘alpha synchronization
bursts’ especially just before the driving errors. The occur-
rence of these bursts was consistent with the nine severe
driving error events (before the events). Alpha synchroni-
zation bursts were also observed before minor driving
errors (in two cases), and additionally two more times inde-
pendently from any driving error. The bursts were more
Fig. 2. The traces of eight-channel EEG, EOG and EMG during 45 s before and 13 s after a severe driving error (Subject No. 4, Date: 2 June 2005). The
row indicates the onset of the driving error, and the green bar the subject’s reaction due to the driving instructor’s correction. We observe the frequent
occurrence of brief 1–2 s paroxysmal bursts of alpha activity in the central and parietal brain regions (orange highlight) with a frequency component
1–2 Hz slower than the alpha rhythm and an increased synchrony among EEG channels (indicated by small arrows between the bursts) that appears more
obvious exactly before the driving error. The alpha bursts just before the error were accompanied with a long-duration eye blink. At the bottom of each
time traces panel, the time-frequency map representation of channel C3 is shown. We observe the increase of alpha power before the driving event that
appears more dominant almost half second before the driving error onset. After the driving instructor’s correction, the alpha rhythm disappears and the
brain activity is shifted to higher frequencies.
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1911
Author's personal copy
prominent in the most sleep-deprived subjects (i.e., subjects
with low MWT scores). As to the wave morphology, they
had frequency typically between the higher portion of the
theta band and the lower portion of alpha bands
(7–9 Hz) with gradual increase of amplitude (Fig. 2). They
can be characterized by high intra- and inter-hemispheric
synchronization among EEG channels. As to the topo-
graphical distribution, these bursts were more dominant
in the central and parietal areas (Fig. 3).
Within this ‘alpha synchronization bursts’ period, eye
blinks of long-duration and sometimes multiple eye blinks
were present (Fig. 2). We also noticed that the same obser-
vations occurred before an increase of EMG activity in cases
where no driving error event was present. This may be inter-
preted as arousal effort of the driver, which could prevent
him/her from the driving error event. During the evaluation
process the two neurophysiologists posed six criteria,
according to their observations, which can characterize the
potential driving error event: (a) at least 10 s before the driv-
ing error event, ‘alpha synchronization bursts’ from the
lower limit of alpha waves gradually to theta waves; (b)
10 s before the driving error event, at least one eye blink
per second and during the driving error event one long-
duration eye blink; (c) decrease of eye blinks occurrence
right after the driving event; (d) after the driving error event,
faster EEG waves and a desynchronization among EEG
channels for about 1–2 s; and (e) increased EMG activity
during the driving error event. Based on the measurements
of the present study, the two neurophysiologists concluded
that provided three of those criteria are met, there is a high
possibility of an upcoming driving error event.
3.2. MWT score
Before the experiment, MWT score did not vary as a
function of driver’s age, or of driver’s experience. There
was not any significant correlation between the MWT score
and any quantitative EEG statistic of the ten first minutes
for any EEG channel. We also did not observe any corre-
lation between the MWT score and the eye blink statistics
for the first 10-min of the measurements.
3.3. Relative frequency band power ratio
The Spectral RBR was significantly increased between
the first and the last 15-min in delta band for the channels
C3, P3, O1, P4, and O2, and the same in alpha band only
for the channel P3. Graphical representations of relative
alpha power suggested that the alpha power increased
between the first and the last 15-min at most sites, but these
results failed to reach significance. This was probably due
to extreme intersubject variations. A statistically significant
decrease between the first and the last 15-min was observed
in beta band for all channels except the frontal one, and in
gamma band for the channels C3, P3, C4, and P4. Theta
RBR did not significantly alter between the first and the
last 15-min.
Single-factor ANOVA did not reveal a statistical signifi-
cant alteration of RBR for theta, alpha, and gamma fre-
quency bands by driving time for any EEG channel. A
statistically significant alteration of RBR in the delta band
by driving time (15-min periods) was revealed for the fol-
lowing EEG channels: C3 (F(3, 116) = 3.006, p= 0.033),
Fig. 3. The absolute power topographical maps for the standard EEG frequency bands before, during, and after the driving error presented in Fig. 2.We
observed a significant increase of alpha and theta power before and during the driving event. Delta power is also increased during the driving error. The
driving error occurred at 3020 s.
1912 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
P3 (F(3, 116) = 5.066, p= 0.002), and O2 (F(3, 116) = 3.147,
p= 0.028). For the beta frequency band, we observed a
statistically significant alteration of RBR by driving time
for the channels: C3 (F(3, 116) = 3.051, p= 0.031), P3
(F(3, 116) = 4.005, p= 0.009), C4 (F(3, 116) = 3.309, p=
0.023), and P4 (F(3, 116) = 3.373, p= 0.021). The post hoc
analysis by Tukey HSD revealed two homogenous groups
for the delta band (channels: C3, P3, and O2), and also two
homogenous groups for the beta band (channels: C3, P3,
C4, and P4), and two for the gamma band (channels: C4
and P4).
3.4. Shannon Entropy, K–L Entropy, and ApEn Measures
A statistically significant decrease between the first and
the last 15-min was observed in Shannon Entropy for all
channels except the frontal one. The statistical analysis for
Shannon Entropy (one-way ANOVA) revealed a statisti-
cally significant alteration by driving time (15-min periods)
for all EEG channels except the frontal ones: C3
(F(3, 116) = 3.77, p= 0.013), P3 (F(3, 116) = 5.846, p=
0.001), O1, (F(3, 116) = 3.758, p= 0.013), C4 (F(3, 116) =
4.451, p= 0.005), P4 (F(3, 116) = 5.521, p= 0.001), and O2
(F(3, 116) = 3.99, p= 0.01).
A statistically significant decrease between the first and
the last 15-min was observed in K–L Entropy for the chan-
nels: P3, O1, P4, and O2. The K–L Entropy was signifi-
cantly affected by driving time only for the P4 channel
(F(3, 116) = 3.51, p= 0.018). The post hoc analysis by
Tukey HSD revealed two homogenous groups for the
Shannon Entropy (all channels except the frontal), and also
two homogenous groups for the K–L Entropy only for the
P4 channel. We did not observe a statistically significant
alteration of ApEn by driving time for any EEG channel.
3.5. Coherence and Cross-ApEn Measures
Coherence increased significantly between the first and
the last ten minutes for the EEG between channels
C3–C4 (F(1, 38) = 5.748, p= 0.022) (Fig. 4). Single-factor
ANOVA revealed a statistically significant alteration of
coherence by driving time (15-min periods) for the same
combination of EEG channels (F(3, 116) = 2.712,
p= 0.048). The statistical analysis for coherence revealed
also a statistically significant increase for the EEG delta
frequency band between the first and the last 10-min
between channels C3–C4 (F(1, 38) = 8.930, p= 0.005) as
well as for the theta frequency band (F(1, 38) = 7.532,
p= 0.009) (Fig. 4). The increase was also observed for
the delta band between channels C3–P3 (F(1, 38) = 6.275,
p= 0.017). ANOVA revealed a statistically significant
alteration of coherence by driving time (15-min periods)
for the channels’ combination C3–C4, for the delta fre-
quency band (F(3, 116) = 5.763, p= 0.001), as well as for
the theta frequency band (F(3, 116) = 5.020, p= 0.003).
By selecting a smaller time period (10-min), single-factor
ANOVA revealed also a statistically significant alteration
of coherence by driving time for the same combination of
channels and the same frequency bands: (delta band:
(F(3,116) = 3.185, p= 0.010) and beta band:
(F(3,116) = 2.827, p= 0.019)).
Single-factor ANOVA revealed a statistically significant
alteration of Cross-ApEn by driving time (15-min periods)
for the following EEG channel combinations: P3–P4,
P3–C4, C3–P4, C4–P4, C3–C4, and C3–P3 (Supplementary
Table 1). For the 10-min periods, significance was reached
only for the EEG channel combination P3–P4
(F(3,116) = 2.544, p= 0.032). Cross-ApEn was increased
significantly between the first and the last quarters for all
the above EEG channel combinations: P3–P4 (p= 0.003),
P3–C4 (p= 0.004), C3–P4 (p= 0.017), C4–P4 (p= 0.019),
C3–C4 (p= 0.02), and C3–P3 (p= 0.033). The post hoc
analysis (15-min periods) revealed two homogenous sub-
sets for these channel combinations.
3.6. Eye blinks
A statistically significant increase between the first and
the last 15-min was observed for the per minute averaged
eye blink duration (PERCLOSE) (F(1,58) = 5.668,
p= 0.021) (Fig. 5). Single-factor ANOVA did not reveal
a statistically significant alteration of per minute averaged
eye blink duration (PERCLOSE) by driving time
(F(3,116) = 2.106, p= 0.103). We did not observe
Fig. 4. The averaged (per 10-min) Coherence for the channels combination C3–C4 for the whole frequency range (left panel), for the delta band (middle
panel), and for theta band (right panel). The p-values correspond to the results of ANOVA, and the asterisks (
*
and
**
) indicate statistically significant
differences (p< 0.05 and p< 0.01) between the first and last 10-min.
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1913
Author's personal copy
significant alteration of number of eye blinks by driving
time (F(3,116) = 0.566, p= 0.638), nor statistically signifi-
cant differences between the first and the last 15-min
(F(1,58) = 0.311, p= 0.579).
The ELS system was found to have high reliability in the
estimation of both the per minute averaged number of eye
blinks (mean ± SD: 96.3% ± 5.2%) and per minute aver-
aged eye blink duration (mean ± SD: 92.8% ± 6.12%).
3.7. Heart rate variability
We did not observe a statistically significant alteration
of heart rate variability by driving time, nor a significant
difference between the first or the last quarters of the
experiment.
3.8. EEG statistics before, during, and after driving error
events
The spectral alpha RBR significantly increased between
the control condition and before the driving errors for the
following channels: C3 (Fig. 6), P3, O1, C4, and P4
(Fig. 7). The increase was more dominant in the central
and parietal channels (Fig. 7). A statistically significant
decrease of gamma RBR was also observed (Fig. 7)
between the control condition and all the segments before
the driving errors for the following channels: Fp1, C3
(Fig. 6), and P3. For the rest of channels, we observed a
statistically significant decrease of gamma RBR between
the control condition and all the segments before and after
the driving errors. For the delta and theta frequency bands,
we did not observe differences between the control condi-
tion and the segments before the driving errors (Fig. 6).
A statistically significant decrease was also observed
between the control condition and all the segments before
the driving errors for beta frequency band, for channel
C4. For the channels Fp1, C3, P3, O1, Fp2, P4, and O2,
we observed a significant decrease between the control con-
dition and some of the segments before the driving errors,
but these results were not consistent for all EEG segments
before the driving event (Fig. 6, for channel C3).
Single-factor ANOVA revealed statistically significant
differences between the 11 EEG segments for the RBR of
theta, alpha, beta, and gamma frequency bands (Supple-
mentary Table 2). The post hoc analysis revealed three
homogenous subsets for alpha band for the channels C3
and C4, including one subset that containing the control
condition and the segments after the driving error, one sub-
set that contained the segment after the driving error (9 and
10) and the segments (1, 2, and 4), and another subset that
contained all the segments before the driving error. Post
hoc analysis revealed three homogenous subsets for theta
band for the channels: Fp1, C3, P3, O1, and C4, and two
homogenous subsets for the rest of channels. For the beta
frequency band, post hoc analysis revealed four homoge-
nous subsets for channels Fp1 and C4, and three subsets
for the other EEG channels. The post hoc analysis revealed
four homogenous subsets for gamma band for the channels
P3, O1, and C4, including one subset that contained the
control condition, one subset that contained the segment
after the driving error (9 and 10), one segment contained
the first segment after the driving event and the first seg-
ment before the driving event (1 and 9), and another subset
that contained all the segments before the driving error.
For the rest of channels, post hoc analysis revealed three
homogenous subsets for the gamma frequency band.
The Shannon Entropy was significantly decreased
between the control condition and most of the segments
before the driving errors for the channels: C3 (Fig. 8),
P3, O1, C4, P4, and O2. For the K–L Entropy, we did
not observe consistent differences between the control con-
dition and the segments before the driving errors (Fig. 8).
Single-factor ANOVA revealed statistically significant dif-
ferences between the 11 EEG segments in all channels for
the Shannon Entropy (Supplementary Table 3), and in all
channels except the frontals for the K–L Entropy (Supple-
mentary Table 3). The post hoc analysis revealed four
homogenous subsets for the Shannon Entropy for channel
C4, with one subset contained the control condition and
the segments after the driving error. For the rest of chan-
nels, post hoc analysis revealed three homogenous subsets
for the Shannon Entropy. For the K–L Entropy, post
Fig. 5. The averaged (per 15-min) number of eye blinks (left panel) and per minute averaged eye blink duration (PERCLOSE) (right panel). The p-values
correspond to the results of ANOVA, and the asterisks (
*
and
**
) indicate statistically significant differences (p< 0.05 and p< 0.01) between the first and
last 15-min.
1914 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
hoc analysis revealed three subsets for channels C4 and P4,
two subsets for channels: C3, P3, O1, and O2, and one sub-
set for the frontal channels. We did not observe statistically
significant differences of Approximate Entropy between the
control condition and the segments before the driving
errors for any channel.
Coherence increased significantly before driving events
in comparison with the control condition for the frequency
band 7–9 Hz (channels combination C3–C4) (Fig. 9), and
for the gamma frequency band (channels’ combination:
C3–C4, C3–P3, C3–P4, C4–P4, P3–C4, and P3–P4).
Single-factor ANOVA revealed statistically significant
Fig. 7. The relative power of alpha and gamma frequency bands of time evolution (in 5-s segments, averaged over nine driving error events) before,
during, and after the driving errors for all eight EEG channels. We observed an increase of alpha and decrease of gamma RBR before the driving error
onset, which is more dominant at the central and parietal channels. The segments correspond to 40 s before (1–8) and 10 s after (9–10) the driving errors’
onset. As the control condition, we selected an EEG segment (5 s duration) from the first minute of each measurement where driving errors occurred.
Fig. 6. The averaged (per 5 s among all nine severe driving events occurred in our experiments) relative band ratio (RBR) of all EEG standard frequency
bands, 40 s before (on red) and 10 s after (on green) the driving errors’ onset. As control condition (on white) we selected EEG segments (5 s duration)
from the first minute of each measurement in which driving errors occurred. The p-values correspond to the results of ANOVA, and the asterisks (
*
and
**
)
indicate statistically significant differences (p< 0.05 and p< 0.01) between the control condition and each time-period.
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1915
Author's personal copy
differences of coherence between the 11 EEG segments for
the alpha band (channels combination C3–C4), and for the
gamma band (channels’ combination: C3–C4, C3–P3, C3–
P4, C4–P4, P3–C4, and P3–P4) (Supplementary Table 4).
The post hoc analysis revealed three homogenous subsets
for coherence in the gamma band (channels’ combination:
C3–C4, C3–P4, C4–P4, P3–C4, and O1–O2).
The Cross-ApEn significantly decreased before the
driving errors in comparison with the control condition
for all the EEG segments for the following channels’
combinations: C3–C4, C3–P3, C3–P4, C4–P4, P3–C4,
P3–P4, and O1–O2 (Fig. 10). Single-factor ANOVA
revealed statistically significant differences between the
11 EEG segments in all previously mentioned channels’
combinations for the Cross-ApEn (Supplementary Table
5). The post hoc analysis revealed three homogenous
subsets for Cross-ApEn for all these channels’
combinations.
Fig. 8. The averaged (per 5 s among all nine severe driving events occurred during our experimental procedure) Shannon Entropy and K–L Entropy of
channel C3, 40 s before (on red) and 10 s after (on green) the driving errors’ onset. As control condition we selected EEG segments (5 s duration – on
white) from the first minute of each measurement in which driving errors occurred. The p-values correspond to the results of ANOVA, and the asterisks (
*
and
**
) indicate statistically significant differences (p< 0.05 and p< 0.01) between the control condition and each time-period. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of this paper.)
Fig. 9. The PLV (7–9 Hz), Coherence (7–9 Hz), and Cross Approximate Entropy between all EEG pairwise channel combinations during the control
condition (first row), before (second row), and after (third row) of a severe driving error.
1916 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
3.9. Eye blink statistics before and after the driving error
events
Both the number of eye blinks and the per minute aver-
aged eye blink duration (PERCLOSE) significantly
increased during the minute right before the driving errors
in comparison with the control condition (number of eye
blinks: F(1, 16) = 5.632, p= 0.030; PERCLOSE:
F(1, 16) = 17.821, p= 0.001) (Fig. 11).
4. Discussion
Driving is a complex task that requires an optimum level
of alertness to guarantee the security of the driver and
other road users. Driver’s sleepiness has been recognized
as one of the most significant safety hazards in the trans-
portation industry (Lal and Craig, 2005). The development
of an on-board countermeasure system that provides driv-
ers with useful feedback about the onset of their sleepiness
can essentially contribute to the prevention of sleep-
deprived traffic accidents.
In the last few decades, there is an increasing interest in
the scientific community on the evaluation of electrophys-
iological indicators for the detection of sleepiness. While
numerous physiological indicators are available for assess-
ing the level of alertness, the EEG signal has been recog-
nized as the most predictive and reliable one (Artaud
et al., 1973; Lal and Craig, 2002). In a number of driving
simulator studies, significant alterations in EEG frequency
bands were observed as a result of fatigue or sleepiness
(Horva
´th et al., 1976; Torsvall and A
˚kerstedt, 1987; Gill-
berg et al., 1996; Lal and Craig, 2000). In general, an
increase in the power of alpha and theta frequency bands
was the consistent finding in the majority of these studies.
Fig. 10. The averaged (per 5 s among all nine severe driving events occurred during our experimental procedure) Cross Approximate Entropy for different
EEG channel combinations, 40 s before (on red) and 10 s after (on green) the driving errors’ onset. As control condition we selected EEG segments (5 s
duration – on white) from the first minute of each measurement in which driving errors occurred. The p-values correspond to the results of ANOVA, and
the asterisks (
*
and
**
) indicate statistically significant differences (p< 0.05 and p< 0.01) between the control condition and each time-period. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.)
Fig. 11. The averaged (per one-minute) number of eye blinks (left panel) and. per minute averaged eye blink duration (PERCLOSE) (right panel) (among
all nine severe driving error events occurred in our experiments) in control condition and before severe driving errors. As control condition we selected
EOG segments (one-minute duration) from the first minute of each measurement in which driving errors occurred. The p-values correspond to the results
of ANOVA, and the asterisks (
*
and
**
) indicate statistically significant differences (p< 0.05 and p< 0.01) between the control condition and the minute
before driving errors.
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1917
Author's personal copy
Notably, we shall not forget that these results have been
obtained in driver simulators, and therefore the subjects
knew that the consequences of their driving errors would
not affect their safety. Moreover, driving simulator pro-
duces results that may not be generalized to real-life driv-
ing, and calibrations have been suggested against real
driving for the driving simulators in various conditions
(Philip et al., 2005).
In the present study, we used an experimental paradigm
that involves neurophysiological measurements in real
driving conditions to assess the driving task in a more real-
istic way. We obtained peripheral physiological measure-
ments and we examined the effectiveness of eye blink
statistics as potential indicators of sleepiness.
It is well known that increased driving time would
impair the alertness and the driving performance (Otmani
et al., 2005). Total sleep restriction during the night prior
to the experiment combined with the effect of time on the
driving task would amplify the decrease in the level of alert-
ness and induce severe sleepiness. Based on this assump-
tion, we estimated the on-going temporal evolution of
physiological measurements during the experiment. We
extended this analysis by also focusing on some
well-defined severe driving errors. Although the idea of
introducing a link between the alertness and driving perfor-
mance is not new (Broukhuis and de Waard, 1993; Cam-
pagne et al., 2004), the recorded neurophysiological
measurements are rarely linked with specific driving errors.
In order to use EEG for the development of a sleepiness
countermeasure device, the EEG signal should be free of
both environmental and biological artifacts that occur fre-
quently in the real field measurements. Artifact rejection
pre-processing techniques were rarely applied to EEG data
recorded from sleep-deprived drivers prior to the quantita-
tive analysis. Gevins et al. (1995), in an operational envi-
ronment experiment, reported advances in on-line
reduction of muscle and eye-movement artifacts based on
low-pass filtering with variable cut-off frequencies. In our
study, we used a widely accepted signal processing tech-
nique, known as independent components analysis (ICA),
in order to remove eye-blinking and muscle activity arti-
facts (Jung et al., 2000). In the current data analysis, we
used the ICA algorithm for off-line EEG analysis; however,
the ICA method can be adapted for real-time processing in
an on-line setup for the development of practical devices.
Our quantitative EEG analysis revealed significant vari-
ations of RBR by driving time in the frequency bands of
delta, alpha, beta, and gamma. These alterations were
more prominent in the central and parietal areas, and less
prominent in the occipital areas. We did not observe
RBR alteration by driving time in frontal electrodes for
any frequency band. More specifically, an increase of
RBR was observed in the delta and alpha bands, and a sig-
nificant decrease in the beta and gamma frequency bands
by driving time. Most of our observations concerning the
EEG frequency bands (increase in delta and alpha bands)
are in agreement with previous findings reported in the lit-
erature. Lal and Craig (2000) found consistent increases in
the delta, theta, and alpha activities during transition to
fatigue from an awake state. Torsvall and A
˚kerstedt
(1987) reported that the alpha activity was clearly the most
sensitive to sleepiness, while the delta and theta bands were
also affected by driving time but to a lesser extent. In a field
study, Kecklund and A
˚kerstedt (1993) recorded EEG con-
tinuously during evening/night driving in a group of truck
drivers, and found increased alpha and theta burst activi-
ties during the last few hours of driving. Although theta
power increase due to sleepiness is a frequent observation
in most of the studies, we did not find a significant alter-
ation of theta RBR as a function of driving time. The
observed beta band RBR decrease is in accordance with
results reported by Lal and Craig (2005), while it is the first
time that the gamma frequency band is investigated, which
has a statistically significant alteration by driving time. The
differences in frequency bands between our results and the
findings of previous studies could be due to the fact that we
estimated RBR instead of the absolute spectral power. In
light of Parseval’s theorem (e.g., Arfken and Weber,
2001), the sum of the square of signal is equal to the sum
of the square of its Fourier transform. Hence, the spectral
power of frequency bands strongly depends on the EEG
signal’s amplitude, and therefore also depends on the elec-
trode’s impedance; while relative power disregards the total
power of the signal, focusing on the ratio of frequency
components of the EEG signal.
Although considerable progress has been made for the
detection of sleepiness onset, we know of no published
studies that analyze the EEG data obtained from sleep-
deprived drivers by using more advanced signal processing
techniques than the standard Fourier analysis. In an
attempt to develop and validate more sensitive and reliable
statistics for sleepiness detection, we also applied several
entropy measures to our EEG data. The concept of
entropy is defined as the rate of information loss over time
(Stam, 2005). The non-linear entropy measures quantify
successfully the cortical function at different sleep stages
(Fell et al., 1996; Acharya et al., 2005). Entropy showed
a very high statistically alteration by driving time in almost
all EEG channels. However, the ApEn was not affected by
driving time, although it seemed to be sensitive to the clas-
sification of sleep stages (Burioka et al., 2005).
Compared to other qualitative descriptive studies, the
finding of Santamaria and Chiappa (1987) unquestionably
provided the most detailed description available for the
EEG events during the sleepiness onset. These authors con-
firmed many earlier observations and stressed that the
EEG changes during drowsiness are often rapid, with some
states lasting from less than a second to a few seconds.
Therefore, the macroscopic temporal event analysis of the
on-going brain activity may not be able to capture these
transient EEG alterations, and important information
might be lost concerning the sleepiness’ onset. For the same
reason, a microscopic event analysis focusing on driving
errors is more desirable.
1918 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
The visual interpretation of our EEG measurements
revealed the occurrence of ‘alpha bursts’ especially right
before the driving errors (Fig. 2). The occurrence of these
bursts was a consistent finding during all measurements
but was more frequent at the end of the measurements,
and more prominent before the errors. Although, these
alpha bursts did not occur exclusively just before the severe
driving errors, the fact that before all nine severe errors
alpha synchronization bursts occurred shows that the
occurrence of brief paroxysmal bursts of alpha activity
and the increased synchrony among EEG channels are
strongly linked with an upcoming driving error. The alpha
bursts all had the same frequency content (in the lower por-
tion of alpha band). These EEG alterations were observed
in previous laboratory sleepiness-onset studies as first
described by Santamaria and Chiappa (1987) and latter
confirmed by Broughton and Hasan (1995) about EEG
alterations during sleepiness onset. Mano et al. (1995),in
their real field driving experiment, observed a diffuse slower
alpha activity when the drivers closed their eyelids as the
driving time went on. Although it seems difficult to identify
if the observed alpha bursts correspond to one of the two
types of alpha activity related with sleepiness as they have
been described by Broughton and Hasan (1995), their
duration (1–2 s) and their spatial distribution lead us to
believe that this phenomenon may correspond to the par-
oxysmal bursts of alpha activity (Broughton and Hasan,
1995). Furthermore, the fact that they occurred more fre-
quently in the EEG recordings of the most-deprived sub-
jects confirms our notion that they are strongly related
with high levels of sleepiness. The spectral RBR of alpha
activity was also significantly increased almost one minute
before the driving errors, and such an increase was more
dominant in the central and parietal channels. However,
the alpha RBR presented statistically significant alteration
by driving time only in one channel in the macroscopic
analysis of our EEG data. This can be explained by the fact
that these paroxysmal bursts of alpha waves occurred in
brief groups of 1–2 s duration, and thus their overall fre-
quency power was not strong enough at all channels to
increase significantly the alpha RBR in the macroscopic
event analysis.
The visual interpretation of our recorded data revealed
an increase of EEG signal synchronization among differ-
ent channels before the driving errors, which led us to
examine if we can quantify this phenomenon by using
both linear and non-linear signal processing techniques.
In the last decade, the notion of EEG synchronization
has attracted the attention of many researchers and it
has led to a whole new range of quantitative EEG mea-
sures as well as a number of emerging applications for
monitoring of sleep (Stam, 2005). In macroscopic analysis,
we calculated the frequency-domain based linear coher-
ence measure and the time-domain based non-linear
Cross-ApEn (Fusheng et al., 2001) in order to quantify
the synchrony among the EEG channels. It was found
that coherence increased significantly between the first
and the last 10-min for the channels’ combination
C3–C4, which corresponds to the central areas of the
brain. In the microscopic analysis of our data, coherence
increased significantly before driving events in comparison
with the control condition for both alpha and gamma fre-
quency bands. The Cross-ApEn was more sensitive in
detecting the sleepiness onset since it significantly
decreased before the severe driving errors for the majority
of EEG channel combinations, and moreover it was sig-
nificantly affected by the driving time for these channel
combinations. In addition, we also used PLV to measure
the short-range (1–2 s) phase synchronization while exam-
ining the pairwise channel synchronization before the
severe driving errors.
Our study provides further evidence that a variety of
quantitative EEG statistics are reliable for the detection
of sleepiness and therefore they can be used as sleepiness
indicators in an on-board countermeasure device. How-
ever, this approach encounters several practical problems.
Even if we assume that a reliable EEG-based sleepiness
indicator is available, there are other technical problems
that need be solved. We should not be remiss of the fact
that the system would work reliably in a real environmen-
tal condition where the EEG signal is highly contaminated
with noise. Even if this problem can be solved by the
usage of some on-line artifact rejection techniques, such
as the ICA, the usability of such a device should be recon-
sidered for more practical than technical reasons. Our
group believes that it will be difficult for the drivers to
place jelly electrodes on their scalp before every driving
session, and especially on hairy areas (such as the central
or parietal regions) that appear to be the most sensitive on
the EEG alterations due to sleepiness. More recently,
advanced electrodes have been developed that allow to
amplify the EEG signal in the scalp-electrode, and some
groups (Lal and Craig, 2001a) also conducted research
on the development of miniaturized dry electrodes. How-
ever, other approaches using peripheral neurophysiologi-
cal measurements, which can be obtained with some
driver-friendly techniques, should also be considered and
studied more intensively.
Eye blinks are sensitive to high visual attention
demands. Fewer and shorter duration blinks are often
associated with situations that require intake of important
information such as reading (Ponder and Kennedy, 1928),
city driving (Lecret and Pottier, 1971), and formation flying
(Wilson et al., 1987). Blink closure duration has also been
shown to be sensitive to high overall task demand. Blink
patterns have also been used to provide information about
the operator’s responses to the environmental stimuli and
thus the situational awareness (Fogarty and Stern, 1989;
Wilson, 1992).
Since Aserinsky and Kleitman (1955) reported changes
in spontaneous eye movement patterns not only in sleep
but also in the hypnagogic state, EOG has been used for
assessing alertness, often together with EEG (Isse et al.,
1978; Hori, 1982; Santamaria and Chiappa, 1987; Ota
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1919
Author's personal copy
et al., 1990; Hyoki et al., 1998). Furthermore, EOG has
been widely applied to clinical evaluations, such as drug
effects on alertness (Shigeta et al., 1993), and arousal levels
in psychiatric disorders (Toyoshima, 1991). However, the
relationship between the EOG and EEG measures had
not been clarified at various stages of alertness until Hyoki
et al. (1993). In their study, a significant correlation
between the number of eye movements and EEG powers
at alpha and beta frequency bands was found (Hyoki
et al., 1993).
Mano et al. (1995) first reported in their real field high-
way driving experiment that blinking of eyelids increased
as the driving time went on; and recently, EOG measure-
ments have been obtained (Lal and Craig, 2005) in a simu-
lator study of driver’s sleepiness. However, there is still a
lack of research on the systematic evaluation of
EOG-based statistics as potential indicators for the driver’s
sleepiness. Indeed, Mano et al. (1995) restricted their
observations on a macroscopic evaluation of the on-going
eye-blinking activity, and Lal and Craig (2005) used the
EOG signal for eye blink artifact identification without
presenting any results concerning the eye-blinking activity.
We analyzed the eye-blinking statistics in both a macro-
scopic and microscopic way. We found a statistically signif-
icant alteration of eye-blinking duration by driving time.
More specifically, the eye-blinking duration and the num-
ber of eye blinks both increased as the driving time went
on, but only eye-blinking duration reached significance.
Moreover, the number of eye blinks and the per minute
averaged eye blink duration (PERCLOSE) statistics were
significantly increased during the minute right before the
driving errors in comparison with the control condition.
Our study presented clear evidence that the eye-blinking
statistics are sensitive to the driver’s sleepiness and they
should be considered in the design of a future sleepiness
detection countermeasure device. Since advanced
technological systems, such as the ELS system used in
our experiment, are available and reliable in measuring
the eye-blinking activity, the EOG-based approach seems
to be more driver-friendly and efficient than an EEG-based
system, and therefore this technology should attract more
attention from the researchers in industry.
In this paper, we evaluated the effectiveness of a pleth-
ora of neurophysiological statistics in detecting driver’s
sleepiness in a real environment experimental paradigm.
We can conclude that the occurrence of brief paroxysmal
bursts of alpha activity (with frequency 1–2 Hz slower than
the conventional alpha rhythm) in the central and parietal
brain regions, as well as an increased synchrony among the
EEG channels, indicates a high level of sleepiness and a
high possibility of an upcoming severe driving error. The
EEG statistics that quantify the signal’s complexity and
synchronization between channels can serve as potential
indicators of driver’s sleepiness. However, we would like
to emphasize that the most important finding of the present
study is the sensitivity of the peripheral physiological mea-
surements, such as the eye blink statistics, to the driver’s
sleepiness. Further research is anticipated for developing
and testing the applicability of a neurophysiological sleep-
iness countermeasure device in real field driving situations,
which should be based on peripheral physiological mea-
surements. It is noteworthy that we have conducted several
multiple tests (each with omnibus hypothesis) on the level
of a single ANOVA and made appropriate corrections of
significance level, a potential remaining pitfall of this pro-
cedure is the existence of falsely claimed significance; in
the literature, it is known the level of false discoveries
can be controlled by false discovery rate (Benjamini and
Hochberg, 1995).
5. Conclusion
The quantitative EEG and EOG statistics are both
promising neurophysiological indicators of sleepiness and
have the potential for monitoring sleepiness in an occupa-
tional setting as well as being used in a sleepiness counter-
measure device. In the present study, we describe in detail
for the first time, the occurrence of brief 1–2 s paroxysmal
bursts of alpha activity (with a frequency concept 1–2 Hz
slower than the conventional alpha rhythm) in the central
and parietal brain regions and an increased synchrony
among the EEG channels before severe driving errors.
We also interpret our findings based on some early pub-
lished neurophysiological studies regarding the sleepiness
onset. The occurrence of these alpha wave bursts indicates
a high level of sleepiness and a high possibility of an
upcoming severe driving error. Finally, we present clear
evidence that eye blink statistics are sensitive to the driver’s
sleepiness and should be considered in the design of a
future sleepiness detection countermeasure device, consid-
ering the fact that advanced technological systems are
now available that can reliably measure eye-blinking activ-
ity. The EOG-based sleepiness detection approach seems to
be more driver-friendly and efficient than an EEG-based
system, and we expect it will attract more attention from
the researchers in the near future.
Acknowledgements
The research was supported by the SENSATION
Project of the Information Society Technologies (IST)
Program (507231) of European Union (EU). We would like
to acknowledge Dr. G. Strikis (IASI Medical Centre, Thes-
saloniki, Greece) for his work on the visual interpretation
of the EEG data. Dr. C. Papadelis would like to thank
his students from the Technological Institute of Thessalo-
niki, Department of Automation, who participated as
subjects in the present study.
Appendix A. Supplementary data
Supplementary data associated with this article can be
found, in the online version, at doi:10.1016/j.clinph.
2007.04.031.
1920 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
Author's personal copy
References
Acharya UR, Faust O, Kannathal N, Chua T, Laxminarayan S. Non-
linear analysis of EEG signals at various sleep stages. Comput
Methods Programs Biomed 2005;80(1):37–45.
A
˚kerstedt T, Torsvall L, Gillberg M. Sleepiness and shiftwork: field
studies. Sleep 1982;5(Suppl. 2):95–106.
Arfken GB, Weber HJ. Mathematical methods for physicists. Har-
court: San Diego; 2001.
Artaud P, Planque S, Lavergne C, Cara H, de Lepine P, Tarriere C, et al.
An on-board system for detecting lapses of alertness in car driving. In:
14th ESV conference, session 2, intelligent vehicle highway system and
human factors, Munich, Germany; 1973.
Aserinsky E, Kleitman N. Two types of ocular motility occurring in sleep.
J Appl Physiol 1955;8:1–10.
Benjamini Y, Hochberg Y. controlling the false discovery rate: a practical
and powerful approach to multiple testing. J R Stat Soc
1995;57(1):289–300.
Bocca ML, Denise P. Total sleep deprivation effect on disengagement of
spatial attention as assessed by saccadic eye movements. Clin
Neurophysiol 2006;117:894–9.
Broughton R, Hasan J. Quantitative topographic electroencephalographic
mapping during drowsiness and sleep onset. J Clin Neurophysiol
1995;12(4):372–86.
Broukhuis KA, de Waard D. The use of psychophysiology to assess driver
status. Ergonomics 1993;36:1099–110.
Burioka N, Miyata M, Cornelissen G, Halberg F, Takeshima T, Kaplan
DT, et al. Approximate entropy in the electroencephalogram during
wake and sleep. Clin EEG Neurosci 2005;36(1):21–4.
Campagne A, Pebayle T, Muzet A. Correlation between driving errors and
vigilance level:influence of the driver’s age.Physiol Behav 2004;80:515–24.
Connor J, Whitlock G, Norton R, Jackson R. The role of driver sleepiness
in car crashes: a systematic review of epidemiological studies. Accid
Anal Prev 2001;33:31–41.
Delorme A, Makeig S. EEGLAB: an open source toolbox for analysis of
single-trial EEG dynamics including independent component analysis.
J Neurosci Methods 2004;134:9–21. (EEGLAB software available
from: http://sccn.ucsd.edu/eeglab).
Dinges DF, Kribbs NB. Performing while sleepy: effects of experimentally
induced sleepiness. In: Monk TH, editor. Sleep, sleepiness and
performance. Chichester: Wiley; 1991. p. 97–128.
Doghramji K, Mitler MM, Sangal RB, et al. A normative study of the
maintenance of wakefulness test (MWT). Electroencephalogr Clin
Neurophysiol 1997;103(5):554–62.
Fell J, Roschke J, Mann K, Schaffner C. Discrimination of sleep stages: a
comparison between spectral and nonlinear EEG measures. Electro-
encephalogr Clin Neurophysiol 1996;98(5):401–10.
Fogarty C, Stern JA. Eye movement and blinks: their relationship to
higher cognitive processes. Int J Psychophysiol 1989;8:35–42.
Fusheng Y, Bo H, Qingyu T. Approximate entropy and its application in
biosignal analysis. In: Akay M, editor. Nonlinear biomedical signal
processing. Dynamic analysis and modeling 2001;vol. II. New Jersey,
USA: IEEE Press Series on Biomedical Engineering; 2001.
Gentili R, Papaxanthis C, Pozzo T. Improvement and generalization of
arm motor performance through motor imagery practice. Neurosci-
ence 2006;137:761–72.
Gevins A, Leong H, Du R, Smith ME, Le J, DuRousseau D, et al.
Toward measurement of brain function in operational environments.
Biol Psychiatry 1995;43:145–52.
Gillberg M, Kecklund G, A
˚kerstedt T. Sleepiness and performance of
professional drivers in a truck simulator – comparisons between day
and night driving. J Sleep Res 1996;5:12–5.
Gray R. Entropy and information theory. New York: Springer; 1990.
Hakkanen H, Summala H. Sleepiness at work among commercial truck
drivers. Sleep 2000;23:49–57.
Harrison Y, Horne JA. The impact of sleep deprivation on decision
making: a review. J Exp Psychol Appl 2000;6(3):236–49.
Hori T. Electrodermal and electro-oculographic activity in a hypnagogic
state. Psychophysiology 1982;19:668–72.
Horne JA, Baulk SD. Awareness of sleepiness when driving. Psychophys-
iology 2004;4:97–110.
Horva
´th M, Frantik E, Kopriva K, Meissner J. EEG theta activity
increase coinciding with performance decrement in a monotonous
task. Activitas Nervosa Superior (Prague) 1976;18:207–10.
Hyoki K, Shigeta M, Nishikawa Y, Kawamuro Y, Soda K. The correlation
between changes of EEG and eye movements (EOG) accompanied by
lowered arousal level. Jpn J Clin Pathol 1993;41:813–8.
Hyoki K, Shigeta M, Tsuno N, Kawamuro Y, Kinoshita T. Quantitative
electro-oculography and electroencephalography as indices of alert-
ness. Electroencephalogr Clin Neurophysiol 1998;106:213–9.
Isse K, Kojima T, Ando H, Shimazono Y, Ando K. A study of the
relationship between eye movements (EOG) and arousal levels (EEG).
Seishin Igaku 1978;20:537–44.
Jung TP, Makeig S, Humphries C, Lee TW, Mckeown MJ, Iraqui V, et al.
Removing electroencephalographic artifacts by blind source separa-
tion. Psychophysiology 2000;37:163–78.
Kecklund G, A
˚kerstedt T. Sleepiness in long distance truck driving: an
ambulatory EEG study of night driving. Ergonomics 1993;36:1007–17.
Khardi S, Vallet M. Drivers vigilance. Analysis of differences in vigilance
states assessments by physiological and mechanical indicators. In:
Proceedings of the 1st World Congress on applications of transport
telematics and intelligent vehicle-highway systems. 30 November–3
December, Paris, France; 1994.
Killgore WD, Balkin TJ, Wesensten NJ. Impaired decision making
following 49 h of sleep deprivation. J Sleep Res 2006;15(1):7–13.
Lachaux J-P, Rodriguez E, Martinerie J, Varela FJ. Measuring phase
synchrony in brain signals. Human Brain Mapp 1999;8:194–208.
Lal SKL, Craig A. Psychophysiological effects associated with drowsiness:
driver fatigue and electroencephalography. Int J Psychophys 2000;35:39.
Lal SKL, Craig A. Electroencephalography activity associated with driver
fatigue: implications for a fatigue countermeasure device. J Psycho-
physiol 2001a;15:183–9.
Lal SKL, Craig A. A critical review of the psychophysiology of driver
fatigue. Biol Psychol 2001b;55:173–94.
Lal SKL, Craig A. Driver fatigue: electrophysiology and psychological
assessment. Psychophysiology 2002;39:1–9.
Lal S, Craig A. Reproducibility of the spectral components of the electroen-
cephalogram during driver fatigue. Int J Psychophysiol 2005;55:137–43.
Lecret F, Pottier M. La vigilance, facteur de securite dans la conduite
automobile. Le Travail Humain 1971;34:51–68.
Le Van Quyen M et al.. Comparison of Hilbert transform and wavelet
methods for the analysis of neuronal synchrony. J Neurosci Methods
2001;111:83–98.
Lin C-T, Wu R-C, Jung T-P, Liang S-F, Huang T-Y. Estimating driving
performance based on EEG spectrum analysis. EURASIP J Appl
Signal Process 2005;19:3165–74.
Mano T, Shiozawa Z, Amino A, Nagasaka T, Shindo K, Nagasaki H,
Kobayashi T, Mihara K. EEG findings for drivers during driving in a
highway. Electroencephalogr Clin Neurophysiol 1995;97(4):149 (1).
Mascord DJ, Heath RA. Behavioral and physiological indices of fatigue in
a visual tracking task. J Safety Res 1992;23:19–25.
O’Hanlon SF, Kelley GR. Comparison of performance and physiological
changes between drivers who perform well and poorly during
prolonged vehicular operation. In: Mackie RR, editor. Vigilance. New
York: Plenum; 1977. p. 189–202.
Ota T, Ohshima H, Toyoshima R, Yamauchi T. Arousal level measure-
ments by biphasic changes of the mean alpha amplitude. Jpn J EEG
EMG 1990;18:258–67.
Otmani S, Pebayle T, Roge J, Muzet A. Effect of driving duration and
partial sleep deprivation on subsequent alertness and performance of
car drivers. Physiol Behav 2005;84:715–24.
Papadelis C, Kourtidou-Papadeli C,Bamidis P, Albani M. Effects of imagery
training on cognitive performance and use of physiological measures
as an assessment tool of mental effort. Brain Cogn 2007;64(1):74–85.
C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922 1921
Author's personal copy
Philip P. Sleepiness of occupational drivers. Ind Health 2005;43:30–3.
Philip P, Sagaspe P, Taillard J, Valtat C, Moore N, A
˚kerstedt T, et al.
Fatigue, sleepiness, and performance in simulated versus real driving
conditions. Sleep 2005;28(12):1511–6.
Pincus SM. Approximate entropy (ApEn) as a complexity measure. Proc
Natl Acad Sci USA 1991;88:110–7.
Ponder E, Kennedy WP. On the act of blinking. Q J Physiol
1928;18:89–110.
Raidy DJ, Scharff LF. Effects of sleep deprivation on auditory and visual
memory tasks. Percept Mot Skills 2005;101(1):451–67.
Santamaria J, Chiappa KH. The EEG of drowsiness in normal subjects. J
Clin Neurophysiol 1987;4:327–82.
Shannon CE. A mathematical theory of communication. Bell Syst Tech J
1948;27:379–423.
Shigeta M, Nishikawa Y, Shimuzu M, Usui M, Hyoki K, Kawamuro Y.
Horizontal component of electro-oculogram as a parameter of arousal
in dementia: relationship between intellectual improvement and
increasing arousal under pharmacotherapy. J Clin Pharmacol
1993;33:741–6.
Stam CJ. Nonlinear dynamical analysis of EEG and MEG: review of an
emerging field. Clin Neurophysiol 2005;116(10):2266–301.
Takens F. Detecting strange attractors in turbulence. In: Rand DA,
Young LS, editors. Dynamical systems and turbulence. Lecture notes
in mathematics 1981;vol. 898. Berlin: Springer; 1981. p. 366–81.
The Parliament of the Commonwealth of Australia. Beyond the midnight oil.
Managing fatigue in transport. House of Representatives Standing
Committee on Communications,Transport and the Arts,Australia; 2000.
Torsvall L, A
˚kerstedt T. Sleepiness and irregular work hours. Sleep Res
1983;12:376.
Torsvall L, A
˚kerstedt T. Sleepiness on the job: continuously measured
EEG changes in train drivers. Electroencephalogr Clin Neurophysiol
1987;66:502–11.
Toyoshima R. Arousal level in schizophrenia: measurements by biphasic
changes of the mean alpha amplitude and frequencies of eye
movements. Jpn J EEG EMG 1991;19:364–73.
Williamson AM, Feyer AM. Moderate sleep deprivation produces impair-
ments in cognitive and motor performance equivalent to legally
prescribed levels of alcohol intoxication. Occup Environ Med
2000;57:649–55.
Wilson GF. Applied use of cardiac and respiratory measures: practical
considerations and precautions. Biological Psychol 1992;34:163–78.
Wilson GF, Purvis B, Skelly J, Fullenkamp P, Davis I. Physiological data
used to measure pilot workload in actual flight and simulator
conditions. In: Proceedings of the human factors society, 31st annual
meeting; 1987, p. 779–83.
Volow MR, Erwin CW. The heart rate variability correlates of sponta-
neous drowsiness onset. International Automotive Engineering Con-
gress; Detroit MI, USA, 1973.
1922 C. Papadelis et al. / Clinical Neurophysiology 118 (2007) 1906–1922
